Inside YC's AI Playbook

Overview

• Speaker: Pete Koomen, Garry Tan, Diana Hu, Harj Taggar, Jared Friedman
• Channel: Y Combinator
• Main topic: Building AI native companies and internal AI infrastructure at Y Combinator
• Purpose: To share YC internal AI playbook publicly for the first time, detailing how AI agents, tool registries, and context layering transform organizational workflows. Y Combinator partners discuss how they have built internal AI infrastructure, agents, harnesses, and shared organisational context tools, transforming YC into an AI-native organization and offering insights into how startups and companies can adopt AI agents for exponential productivity.

Topic Map

YC's Internal AI Infrastructure & Agent Adoption

• Explanation: Pete Koomen details how YC started building internal AI agent harnesses, tool registries, and database integrations for finance and partner workflows over the past year.
• Key claims:
  • YC built internal agent infrastructure and harnesses to empower every team.
  • A single Postgres database containing all company, founder, and financial context powers YC agents.
  • Agent tool registries enable teams to add custom skills and tools easily.
• Examples:
  • Giving finance agents access to financial workflows, journal entries, and priced rounds.
  • Allowing agents to query database tables and model files directly.
• Terminology:
  • agent loop
  • tool registry
  • shared organizational context
  • model router
• Why it matters: Demonstrates how centralized context and agent harnesses eliminate friction and back-and-forth between teams, supercharging organizational speed.

Horseless Carriages & AI Software Design

• Explanation: Discussion on how early AI software mimics old paradigms by hiding prompts and AI features behind traditional deterministic UIs instead of letting users interact directly with agents.
• Key claims:
  • Many AI apps are 'horseless carriages'—mimicking old software paradigms by locking prompt control behind traditional UI.
  • Shift control of software development and execution from the developer to the user through AI agents.
  • True AI-native software wraps deterministic software tools inside agent wrappers rather than wrapping AI inside deterministic code.
• Examples:
  • Gmail AI assistant hiding prompt context instead of letting users customize or interact freely.
  • Trewittick's Steam Carriage design mimicking horse carriages.
• Terminology:
  • horseless carriages
  • software-deterministic
  • agentic tools
• Why it matters: Explains why many current AI integrations fail to deliver true leverage and outlines the architecture of future AI software.

Skillification, Denormalization, and Two-Sentence Pitches

• Explanation: The partners discuss 'skillification'—converting human workflows and company context into reusable agent skills, such as two-sentence company descriptions.
• Key claims:
  • De-normalizing data and putting all organizational context into a single database enables agents to answer complex arbitrary queries.
  • Two-sentence descriptions act as atomic foundational prompts that every partner and agent uses to understand companies.
  • Multi-player harnesses and team skill sharing represent the next frontier of organizational AI.
• Examples:
  • Tom Blomfield writing custom prompts to condense company context into two sentences.
  • Using MCP (Model Context Protocol) and shared skill registries across open-source tools like OpenClaw and Claude Code.
• Terminology:
  • skillify
  • de-normalization
  • two-sentence description
  • meta-skill
  • check-resolvable
• Why it matters: Shows how standardizing and compressing organizational context into reusable agent skills drastically improves output quality and team onboarding.

Centralization vs. Decentralization in AI Organizations

• Explanation: Examining whether AI leads to centralized control or decentralized employee empowerment, and why transparent default-public agent conversations foster trust.
• Key claims:
  • Default-public agent conversations broadcasted to Slack create internal transparency and social control.
  • AI eliminates the fear of asking 'dumb questions', allowing junior employees to ramp up much faster.
  • True AI-native organizations require being egalitarian and trust-by-default, which runs counter to traditional corporate structures.
• Examples:
  • Broadcasting all agent conversations internally to Slack channels for collaborative learning.
  • New employees using open-source agent instances to query historical context and codebases without bothering senior partners.
• Terminology:
  • social control
  • trust by default
  • egalitarian infrastructure
• Why it matters: Highlights the cultural and structural prerequisites needed for organizations to successfully adopt AI agents at scale.

Key Points

Single Database Context Layer

• Explanation: Storing all company, founder, and financial records in one Postgres database allows agents to execute complex cross-domain queries instantly.
• Evidence: YC's internal database houses company tables, founder tables, financial transactions, and CRM notes in one schema.
• Practical implication: Eliminates departmental silos and data fragmentation, empowering any agent or employee to query comprehensive company context.

The Shift from Coder to User

• Explanation: AI shifts the locus of software control from developers writing deterministic workflows to users prompting and shaping agent behavior directly.
• Evidence: Building custom Rails apps in days versus using agentic tools like Claude Code and OpenClaw to iterate dynamically.
• Practical implication: Reduces software development cycles from months to days, making custom internal tools virtually free to build.

Skillification of Organizational Knowledge

• Explanation: Translating human best practices into reusable agent skills (like two-sentence company descriptions) creates a shared organizational brain.
• Evidence: Partners practicing and refining two-sentence company descriptions until agents internalize the standard.
• Practical implication: Scales expert-level partner knowledge across the entire organization instantaneously.

Frameworks, Models & Processes

The Agentic Harness Framework

• How it works: Combine a single-source database schema, a tool/skill registry, and a model router into an agentic loop that executes domain-specific tasks.
• Components:
  • Agent Loop
  • Tool Registry / Resolver
  • Single Postgres Database Schema
  • Skillification layer
• When to use: When building internal AI infrastructure for an organization to manage workflows, CRM, and financial analysis.

Examples & Case Studies

Building a half-million line Rails app for YC's directory replaced by building a dynamic agentic framework using OpenClaw and Claude Code in days.

• Illustrates: The speed and efficiency of AI-native software development compared to legacy monolithic stacks.
• Lesson: Do not rebuild complex traditional software; wrap AI around modular open-source agent primitives.

Actionable Takeaways

• Immediate:
  • De-normalize and centralize your company's core data into an easily queryable database for AI agents.
  • Start building internal tool registries and agent harnesses for simple workflows like journaling and CRM notes.
  • Practice condensing your product or company description into a precise two-sentence pitch.
• Strategic:
  • Shift software architecture from deterministic code wrapping AI features to AI agents wrapping deterministic tools.
  • Adopt default-public agent conversation channels to foster internal transparency, learning, and social alignment.
  • Embrace egalitarian, trust-by-default organizational structures to unlock the full potential of AI leverage.
• Questions to investigate:
  • How do we balance internal agent transparency with privacy and security for sensitive data?
  • What is the optimal path for legacy enterprises with siloed data to transition to an AI-native architecture?

Claims Worth Verifying

• YC built internal agent infrastructure and tools over the past year that power financial workflows and partner operations. (empirical)
• Default-public broadcasting of agent conversations on Slack improves organizational learning and security through social control. (theoretical)

Notable Quotes

"How do you build super intelligence inside a company?" (at 0:00) "It's like a shared organizational brain. It's like the closest thing to us being able to like connect our brains." (at 0:16) "Whenever a new technology is invented, the first tools built with it inevitably fail because they mimic the old way of doing things." (at 53:47)

Compressed Summary

• YC built internal AI infrastructure centered on agent loops and tool registries.
• Centralizing organizational context in a single Postgres database empowers AI agents to answer complex queries.
• AI shifts control from developers to users through agentic software wrappers.
• Skillification and two-sentence descriptions scale partner expertise across the organization.
• Default-public agent logs foster internal transparency, trust, and rapid employee onboarding.
• Keywords: agents, harness, tool-registry, skillification, context-layer
• Core insight: Building AI-native companies requires centralizing organizational context into a single database and empowering agents with tool registries to shift software control from developers to users.